1288. Feasibility of Fully-Automated Abdominal MRI Organ Segmentation and Volumetry Using a Multi-Class Convolutional Neural Network
Authors * Denotes Presenting Author
  1. Sophie You *; University of California, San Diego
  2. Sulan Wu; California Institute of Technology
  3. Evan Masutani; University of California, San Diego
  4. Brian Hurt; University of California, San Diego
  5. Meagan Rubel; University of California, San Diego
  6. Joy Liau; University of California, San Diego
  7. Albert Hsiao; University of California, San Diego
Abdominal magnetic resonance imaging (MRI) is essential for the evaluation of many medical conditions, for which organ segmentation and volumetry can aid in diagnosis, prognosis, and monitoring. However, such measurements are time-consuming and thus are not routinely performed. While convolutional neural networks (CNNs) have previously been developed for segmentation of individual organs, we considered the possibility that CNNs might be further taught to identify and segment multiple organs simultaneously. Specifically, we investigated the feasibility of using transfer learning to train a single CNN, previously taught to localize and segment the liver on axial images, to automatically identify and segment the liver, spleen, and both kidneys on coronal images. These segmentations might then be used for further organ characterization, including quantification of organ volume.

Materials and Methods:
With Institutional Review Board (IRB) approval and Health Insurance Portability and Accountability Act (HIPAA) compliance, we retrospectively collected a convenience sample of 310 coronal image series from abdominal MRI examinations performed at our institution, including 151 steady-state free precession (SSFP), 128 single-shot fast spin echo (SSFSE), and 31 other pulse sequences. Manual segmentations were supervised by a board-certified radiologist using in-house software. After setting aside 20 SSFP and 20 SSFSE series for testing, the remaining data was divided into 70% for training and 30% for validation. We analyzed segmentation accuracy by calculating Dice similarity between CNN segmentations and ground truth and report its mean. For organ volumetry, we assessed agreement between automated and manual measurements with Pearson correlation and Bland-Altman analysis.

The following results are reported in order of liver, spleen, right kidney, and left kidney. For the validation data, Dice scores were 0.75, 0.75, 0.84, and 0.85. Pearson coefficients were 0.91, 0.73, 0.63, and 0.82. Overall mean difference and 95% limits of agreement (LoA) were 67.68 (-244.94, 380.29) mL. For the 20 SSFP series, Dice scores were 0.87, 0.84, 0.9, and 0.89. Pearson coefficients were 0.91, 0.99, 0.88, and 0.66. Mean difference and 95% LoA were 48.02 (-146.91, 242.95) mL. For the 20 SSFSE series, Dice scores were 0.83, 0.78, 0.88, and 0.84. Pearson coefficients were 0.9, 0.65, 0.67, and 0.54. Mean difference and 95% LoA were 63.29 (-100.99, 227.56) mL.

CNNs have the potential to automate organ segmentation and volumetry on abdominal MRI, aiding in the detection and diagnosis of disease. We show the feasibility of retasking a CNN to segment multiple abdominal organs concurrently in the coronal plane. CNN performance on SSFP images exceeded SSFSE, suggesting that organ signal characteristics on certain pulse sequences may be more favorable for this task. With further refinement, automated segmentations may be used for organ volumetry, as demonstrated here, or other aspects of organ biometry including characterization of MRI tissue properties, perfusion defects, or surface nodularity.